You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[Common] Avoid zero-initialising the NN TPC PID prediction buffer
The per-DataFrame prediction buffer (masked tracks x output dims x 9 mass
hypotheses, ~72 MB for a 1M-track PbPb DataFrame) is a value-initialised
std::vector: a full memset touching every page, immediately overwritten in
full by the per-hypothesis evaluation loop. Carry it as an uninitialised
unique_ptr<float[]> instead, so each page is touched once, by the write that
fills it. All consumers only read it under useNetworkCorrection, so the null
buffer with the network disabled is never dereferenced. No change to any
computed value.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DgKXVM9Q9Hcgexfyj7Eqpp
0 commit comments